Hybrid Neural Network PFNN_FG And Its Application In Plant Disease Prediction

Feedforward neural network(FNN) is the mostly used network. It uses error backpropagation (Bp) training. The backpropagation learning is too slow for many applications and it scales up poorly as tasks become larger and more complex.This thesis aims at investigating the methods of improving the performance of artificial neural networks. Based on the parametric feedforward neural network(PFNN), a new concept of PFNN is redefined, which means an FNN with 2-parameter variant Sigmoid functions for all hidden layers and linear function for output layer, with all layers short-cut connected. The structure of PFNN is modified and fuzzy input layer is added to fit the application of PFNN in plant disease prediction. The genetic algorithm(GA) is applied to configure the parameters of variant Sigmoid function and fuzzy set parameters. Therefore a new network named Parametric Feedforward Neural Network with Fuzzy inputs configured by a Genetic algorithm (PFNNFG) is proposed.On the other hand, experiments with benchmark problems on PFNNFG are done and the performance of PFNNFG is tested. An application of PFNNFG in plant disease prediction is introduced.The first chapter of this thesis gives the basic knowledge of neural networks and the research on neural network as well as the problems of recently used FNNs.The second chapter deals with the foundation of Computational Intelligence (CI). The fundamental of artificial neural network (ANN), Fuzzy theory and Genetic Algorithm are discussed in this chapter.The third chapter of this thesis gives the important ingredients of PFNN_FG: the parametric feedforward neural network, fuzzy sets at the input and output layer and the chromosomal representation of the variant Sigmoid function parameters and fuzzy set parameters as well as the genetic operation. The performance of ANN for plant disease prediction is improved by combining fuzzy set theory and GA. The parametric feedforward neural network using 2-parameter variant Sigmoid functions at all hidden layers improves the learning speed by enlarging the error signals when the signals propagate backwards through the network. The parameters of variant Sigmoid functions (parameter for the dynamic range of the function, for the slope of the function) are configured by GA. The use of only 2 parameters instead of 3 parameters reduces the length of the chromosome string without losing the function to increase the error signals while they propagate backwards. We have shown that using a variant Sigmoid function to fit the transfer curve is an effective means of improving the learning speed of the ANN. The interconnections influence theperformance of the neural network. In our experiments we used the shortcut connections, which also decrease the probability of oscillations, and improve the ability to deal with both linear and nonlinear problems. In order to process linguistic inputs such as: temperature is "high", relative humidity is "low" or "too much" irrigation, fuzzy set theory is employed at the input layer of the neural network. The parameters of membership function are configured by GA. They are encoded in one chromosome as binary string.The 5 benchmark problems (Cancer, Diabetes, Glass, Soybean, and Thyroid problems) are tested with PFNNJFG in the fourth chapter. Performance of PFNNJFG versus FNN, radial basis function(RBF) network, cascade correIation(CC)/pruned cascade correlation(PCC) is tested with all the 5 benchmark problems. Results show that PFNN_FG is superior to other neural networks.Plant disease prediction (PDP) is an important aspect of agricultural. Due to its complexity, pronounced nonlinearity, interactions and multi-variable property, the prediction process is a very challenging task for plant protection. In the fifth chapter, the hybrid neural network is applied to predict plant disease. Three cucumber downy mildew data are tested and the mean square errors(MSEs) are given.The last chapter of this thesis gives a summary of the whole work and presents a system design for plant protection. The realization of this system is t...